Dots can make purchases using cards you’ve already saved on a merchant’s website. Those purchases require your approval (OpenAI, 29 September 2026). That rule makes OpenAI dots a business question and not only a ChatGPT novelty. This guide covers how a dot works, which plan gives access, what a Pro or Business Premium account costs, where the limits sit and what a personal agent with a cloud computer changes for your brand.
What are OpenAI dots and how do they work?
What is an OpenAI dot?
An OpenAI dot is a personal artificial intelligence agent inside ChatGPT that keeps working in the background when the conversation is closed. OpenAI announced OpenAI dots on 29 September 2026 for a ChatGPT account on the Pro or Business Premium plan, with enterprise access in beta. Each dot runs on GPT-6 Astra, the model behind its answers. It learns your priorities. It can read data in your connected apps under the security and privacy terms of your plan. It is an always-on agent rather than a chat window: language models still power each reply; the dot tracks progress, works out what needs to happen next and follows through as things change (OpenAI, learn.chatgpt.com). You give it a responsibility and it owns it. A dot can continue that work for days and store internal notes along the way.
A cloud computer, 4,000 apps and optional access to yours
Every dot has its own cloud computer. Powered by GPT-6 Astra, it can browse websites, read a page, open a file and use an app without touching your device. It connects to more than 4,000 apps through the ChatGPT plugin ecosystem. Connect your calendar, email or Slack application once and the dot can read and act inside them. A single task can span several apps and the dot keeps each task moving. Permissions work app by app: you choose which files, messages and tools a dot may read. Check which permissions each connected app grants before you start. You can also give a dot access to your own computer through the ChatGPT desktop app. That desktop access is optional and stays under your control, with local files and local tools staying yours (OpenAI Help Center). Workspace content from Business, Enterprise and Edu is not used as training data by default (OpenAI, Introducing dots).
Where you reach your dot: ChatGPT, Slack, Teams and text
You reach a dot in Slack, in Microsoft Teams and in ChatGPT, including its desktop app. You can also message or call it from your mobile phone and you choose when and where it contacts you. Mobile access means a task can move forward while you are away from your desk. In a team workspace the dot sits next to your colleagues and can send an update on each task. It can text you when research is done or a step needs your sign-off. Its replies arrive as a message in the channel you chose. You can ask for a status update whenever you like and each update lands where you decided. Stopping those texts is a setting and not a support ticket (OpenAI Help Center).
Memory, proactive research and scheduled work
A dot keeps memory between conversations so you do not restate context every time. It can run proactive research in read-only mode and bring the findings back. You can schedule a recurring task such as a weekly price check or a deadline reminder. Feedback shapes the ongoing work: tell the dot what to change and it adjusts how it handles that task. You can also create new rules for the ongoing responsibilities you hand over. OpenAI’s own example is a tester whose dot spotted a forgotten invoice, prepared it and sent it after confirmation (OpenAI, Introducing dots). It is the vendor’s illustration rather than an independent measurement. OpenAI says it does not train directly on proactive research or on a dot’s internal notes. Read the privacy policy of your plan before you connect a work account. The dot logs its progress, so you can check the details of each step and keep the workflow moving. The Help Center devotes a section to when a dot can start work without a new message from you.
How do dots differ from ChatGPT, Codex and ChatGPT Work?
A ChatGPT conversation ends when you close it. A dot does not: it can continue a task across days, continuing from the context it keeps. Codex is the coding agent for code tasks. It and ChatGPT Work remain separate task spaces. The Help Center explains how a dot uses Codex cloud environments. Tasks launched there count toward your usage limits while chatting with the dot does not (OpenAI, Introducing dots). Dots are a feature of the ChatGPT application. They are not a standalone piece of agentic software. DataCamp reports that they have no public API. Forbes compares dots with Claude Cowork and other agents and concludes that persistence, not smarter replies, is the real advantage (Forbes, 5 October 2026).
Anatomy of a dot-assisted purchase
Follow one purchase from start to finish. A shopper asks a dot to replace a worn-out winter coat. The dot moves through six steps.
First the user states the goal and sets a Custom Rule on spending. Then the dot runs read-only proactive research, including product pages. Third it navigates the merchant’s site in its cloud browser unless the site blocks it. Fourth the auto-review checks any action that touches an account while the dot tracks its progress. Fifth a purchase on a card already saved at the merchant waits for the user’s go-ahead. Last the dot reports what it did and hands sensitive steps to the human. Your brand controls steps two, three and six: readable product data, an open site and exact replies on price and availability.
Who can use OpenAI dots and what do they cost?
Which plans include dots?
Dots come with ChatGPT Pro and Business Premium. Pro is open to users aged 18 and over outside the European Economic Area, Switzerland and the United Kingdom. Business Premium follows the supported ChatGPT regions, for eligible workspaces only. Enterprise, Edu and Healthcare plans get dots in beta, switched off by default until an admin enables them (OpenAI Help Center, consulted 26 October 2026). Plan rules can move so check the details on the page before you budget and schedule any rollout.
Pricing: Pro from $100 a month, with three usage tiers
The first dot is included at no extra cost in the two plans. Pro is listed From $100 a month with three usage tiers, according to the ChatGPT pricing page consulted on 26 October 2026. Additional dots are not on sale yet. Some press coverage quotes $200 a month, a figure that does not match the entry price on that page.
Why is Pro closed in the EEA, Switzerland and the UK?
The plan is not open in the EEA, Switzerland or the UK at launch and OpenAI gives no date. An individual in Paris or Zurich cannot take out Pro for a dot today. Business Premium remains available in supported ChatGPT regions. Some reports tie the exclusion to European regulation. We could not confirm that on a primary source so treat it as unverified.
Do dots count toward ChatGPT usage limits?
Conversations with a dot do not count toward your ChatGPT usage limits. Tasks the dot launches in Codex or ChatGPT Work do count, so heavy tasks can use your allowance (OpenAI, Introducing dots). A light dot adds nothing to your allowance while heavy delegated work still draws on your plan.
How much control do OpenAI dots really give you?
Custom Rules: four behaviors for every action
Custom Rules let you set one of four behaviors for each kind of action: take action without asking, take action if pre-approved, ask before taking action or hand off to you. You can write rules about what a dot may share, buy or look up. These permissions decide how much room a dot has before it asks you (OpenAI Help Center).
Auto-review and approvals
The auto-review step evaluates actions that touch an account before the dot performs them. Purchases still need your explicit approval. You can also pause a dot or reset it, which works as a stop button if a task drifts (OpenAI, Introducing dots). Between steps you can review progress, give feedback and set new permissions. A confirmed action can still be blocked when a site refuses it. The Help Center covers that case in its troubleshooting section.
Sensitive steps a dot hands back to you
Some steps never stay with the agent. Changing a password and transferring money between financial accounts go back to you (OpenAI, How we build safety, security and privacy into dots). For a brand this sets a clear boundary. A dot can research and propose but the most sensitive moves remain human. You can review in the conversation what it did and read each action it took, message by message.
Prompt injection and sites that block cloud browsers
A web page can contain instructions written to hijack an agent. OpenAI calls this prompt injection and builds protections against it. A second constraint is more practical for sellers: some websites block cloud browsers or ask for extra verification. The user must then switch the dot to their own computer with no session transfer and no guarantee the site works (OpenAI Help Center). If your anti-bot rules shut out every cloud-based visitor a dot may never reach your checkout.
Specialist dots for organizations
OpenAI also describes specialist dots for organizations. Internal tests covered purchasing, invoice processing, email marketing, customer support and commercial contracts. Each specialist dot gets its own identity in the company with its own permissions. A team can then treat it as a service with an owner. Specialist dots are coming to Microsoft Agent 365 with targeted enterprise pilots (OpenAI, Introducing dots). A procurement dot would meet your sales site the way a consumer dot does.
What OpenAI dots mean for brands
AI-referred traffic is already worth having
Agent visits are no longer hypothetical. Adobe Digital Insights reports that AI-referred traffic to US retail sites grew +393 % year on year in the first quarter of 2026. Adobe counts visits referred by artificial intelligence assistants. In March 2026 it converted 42 % better than non-AI traffic (Adobe, 16 April 2026, more than one trillion visits). The same study found that product pages are machine-readable only 66 % of the time on average. Home pages reach 75 % and category pages 74 %. These are US retail figures and they show correlation and not causation. They still point to a gap: agents arrive and your product pages are the least readable of the three page types Adobe measured. A language model reading a product page sees text, prices and options. The model cannot guess a product detail the page does not state. A size selector hidden in a script gives it nothing to work with.
Is your site ready for a dot?
Answer ten questions about your own site. Eight cover what a dot does on a merchant site: browsing with a cloud browser, being blocked, paying with a saved card, waiting for confirmation, handing sensitive steps to a human and reading product pages with a language model. Two cover your data: whether your catalog is exposed through a protocol such as MCP and whether product attributes are machine-readable.
The result is indicative and not an audit. It shows which of three levels your site sits at and which three fixes to start with. Use it as a first input for the decision, before a deeper review with your e-commerce leader, a data engineer and a front-end engineer, including the person who manages bot protection. The leader who owns the decision should read the three fixes first. Ask the team that owns your checkout and the team that owns your catalog to answer together. Write down the requirements and the options before you create any project plan. List the checkout software and the catalog software involved. Decide who owns the approval workflow for purchases. Keep the details in one document for your company and share it with the team. Buyers will trust an agent only when a purchase completes without surprises. Budget the analysis and the resources now.
Personal agents vs brand agents
A dot works for the buyer. A brand agent works for the seller. The table sets both side by side on six points.
A dot reads your site as any visitor would. A brand agent responds from data you control. Both personal agents and brand agents will meet in the same conversation. Agents will continue to read your product pages whether you prepare for them or not. Treat readiness as ongoing work and set a review date. Press coverage frames dots as OpenAI’s answer to Meta’s Muse. Meta launched that personal agent on 8 September 2026, according to press reports. We have not checked Meta’s own documentation so we do not compare Muse with dots here. The same readability question applies to Muse and to Claude Cowork, the agents the press sets next to dots. This article relies on primary OpenAI sources wherever they exist.
Three fixes to start with and when to act
One person’s dot is not a threat to your sales. It serves buyers who already know their goal on a plan reserved for paying subscribers and closed to Pro users in Europe. The longer view is different. McKinsey estimates that agentic commerce could orchestrate $900 billion to $1 trillion of US B2C retail revenue by 2030 and $3 to $5 trillion worldwide for goods only (McKinsey, 17 October 2025). Start with three fixes:
- Make product data readable by a machine, product by product, starting with attributes, prices and availability.
- Check that your site does not shut out legitimate cloud visitors.
- Give agents exact replies on price and stock from one governed source. Keep a service owner accountable for that source.
Kleio: making complex catalogs readable to AI agents
A Knowledge Engine for exact answers
An agent that questions a brand needs exact replies and not plausible ones. The Knowledge Engine of Kleio splits catalog data into three purpose-built stores for product, document and memory, with a separate governance layer. It reaches 99 % precision on queries and responds in under three seconds with zero hallucination. We solve the data problem before we deploy any AI agent. A catalog with attributes, prices and constraints can then serve a machine as reliably as an advisor. Our explainer on the Knowledge Engine shows the architecture.
MCP and UCP: exposing your catalog to AI agents
The third surface of Kleio, agent-facing AI search, exposes your product catalog to AI agents and AI search engines such as ChatGPT and Gemini from Google through MCP and UCP. High-intent recommendations reach the buyer where the search starts. Kleio has announced no integration with OpenAI dots. The principle holds for any agent that queries a brand: governed data, one access point and exact replies. Read how MCP and A2A work for enterprise AI search and what ChatGPT Shopping already does on the retail side.
Orpi, Havas Voyages and Selectour: journeys already in production
Orpi put its platform live in three months on Orpi.com and across its 1,250 agencies for 8,000 advisors. Havas Voyages and Selectour run the same approach in travel. Kleio takes a project from kickoff to production in 8 to 12 weeks with seven pre-built industry ontologies and a SOC 2, tenant-isolated setup.
"In three months, we kept our commitment: the platform is live, on Orpi.com and in each of our 1,250 agencies." Guillaume Martinaud, CEO, Orpi
